Research map / Point clouds / Shape and surface completion
SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SymmCompletion, a point cloud completion method that estimates point-wise local symmetry transformations to build partial-missing pairs and then refines them with a symmetry-guided transformer.
Problem
Point cloud completion methods often lose original geometric details and produce geometric inconsistency between the observed partial input and the reconstructed missing regions.
Approach
Replace global symmetry transforms with point-wise affine and translation transforms predicted from local and global features (LSTNet) to generate missing parts, then use features of the partial and missing parts as explicit symmetry guidance via dual-path cross/self-attention refinement (SGFormer).
Key results
- On PCN, SymmCompletion reaches average CD 6.28 (x10^3) and F1 0.853, beating SVDFormer (6.54, 0.841) and CRA-PCN (6.39).
- On MVP at 16384 points, CD 2.14 and F1 0.850, outperforming FBNet (2.29, 0.822).
- On KITTI, it does not achieve the best FD (2.54 vs. SeedFormer 1.45) or MMD (1.72 vs. SVDFormer 0.97), which the authors attribute to domain gap and metric issues.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add both key-geometry features Fk and missing-part features Fm as SGFormer guidance (ablation) | SymmCompletion · PCN | CD (L1 × 10³): 6.51 -> 6.28 | Helped |
| Add both key-geometry features Fk and missing-part features Fm as SGFormer guidance (ablation) | SymmCompletion · PCN | F1-score@%1: 0.839 -> 0.853 | Helped |
| Use local point-wise LSTNet transformations instead of global symmetry transformations (ablation) | SymmCompletion · PCN | CD (L1 × 10³): 6.51 -> 6.28; 6.48 -> 6.28 | Helped |
| Use local point-wise LSTNet transformations instead of global symmetry transformations (ablation) | SymmCompletion · PCN | F1-score@%1: 0.839 -> 0.853; 0.842 -> 0.853 | Helped |
| Use both local features and global awareness to estimate local symmetry (ablation) | LSTNet · PCN | CD (L1 × 10³): 6.41 -> 6.28; 6.49 -> 6.28 | Helped |
| Use both local features and global awareness to estimate local symmetry (ablation) | LSTNet · PCN | F1-score@%1: 0.846 -> 0.853; 0.842 -> 0.853 | Helped |
| Replace the initial point generator with LSTNet (ablation) | SnowflakeNet · PCN | CD (L1 × 10³): 7.21 -> 6.92 | Helped |
| Replace the initial point generator with LSTNet (ablation) | SVDFormer · PCN | CD (L1 × 10³): 6.54 -> 6.47 | Helped |
| Use SGFormer instead of the SPD refinement module (ablation) | SymmCompletion · PCN | CD (L1 × 10³): 7.15 -> 6.28 | Helped |
| Use SGFormer instead of the UpFormer refinement module (ablation) | SymmCompletion · PCN | CD (L1 × 10³): 6.87 -> 6.28 | Helped |
| Use SGFormer instead of the SDG refinement module (ablation) | SymmCompletion · PCN | CD (L1 × 10³): 6.49 -> 6.28 | Helped |
| Use SGFormer instead of the SPD refinement module (ablation) | SymmCompletion · PCN | F1-score@%1: 0.799 -> 0.853 | Helped |
Surprising findings
- Best FD on KITTI is achieved by PoinTr at 0.0 simply by concatenating partial inputs into outputs, which the authors argue shows FD's limitation.
- Global-feature symmetry (USSPA, STN3D) baseline gaps are small on PCN (6.51 and 6.48 vs 6.28 CD).
Limitations
- Struggles to reconstruct true appearance of objects lacking symmetry in challenging inputs.
- Quantitative KITTI results are not best on FD and MMD; real-world evaluation relies on PCN-trained model with metric issues.
Experimental setup
Datasets: PCN MVP ShapeNet55/34 KITTI
Benchmarks: PCN MVP ShapeNet55/34 KITTI
Metrics: Chamfer Distance (L1 and L2) F1-score@1% Fidelity Distance Minimal Matching Distance FLOPs Params inference time
Baselines: PCN PoinTr SnowflakeNet FBNet SeedFormer SVDFormer GTNet AnchorFormer CRA-PCN CRN VRCNet SDNet
Builds on: PCN PoinTr SeedFormer SVDFormer USSPA
Closest papers by meaning
- Point Cloud Completion by Skip-Attention Network With Hierarchical Folding (CVPR 2020 · similarity 0.89)
- Rethinking Multimodal Point Cloud Completion: A Completion-by-Correction Perspective (AAAI 2026 · similarity 0.89)
- Orthogonal Dictionary Guided Shape Completion Network for Point Cloud (AAAI 2024 · similarity 0.88)
- Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.87)
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.87)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.